UniGenDet: A Unified Generative-Discriminative Framework for Co-Evolving Image Generation and Detection

Forum topic · 小凯 · 2026-04-27

Summary

UniGenDet is a unified generative-discriminative framework that enables image generation and AI-generated image detection to co-evolve, addressing the architectural divide between the two fields. Published on arXiv (2604.21904) by researchers including Yanran Zhang, Jiwen Lu, and Jie Zhou, the paper notes that image generation and detection have developed independently: generation relies on generative networks while detection favors discriminative architectures. To bridge this task gap, the authors design a symbiotic multimodal self-attention mechanism and a unified fine-tuning algorithm, alongside a detector-informed generation alignment mechanism for seamless information exchange. This synergy lets the generation task improve interpretability of authenticity recognition, while authenticity criteria guide the creation of higher-fidelity images. Experiments across multiple datasets show state-of-the-art performance for both generation and detection.

Paper Overview

Field: Computer Vision (CV) Authors: Yanran Zhang, Wenzhao Zheng, Yifei Li, Bingyao Yu, Yu Zheng, Lei Chen, Jiwen Lu, Jie Zhou Published: 2026-04-23 arXiv: 2604.21904

Abstract

In recent years, both image generation and AI-generated image detection have made significant progress. Despite their rapid development, the two fields have largely evolved independently, forming distinct architectural paradigms: the former relies primarily on generative networks, while the latter favors discriminative frameworks. A recent trend in both areas is the use of adversarial information to improve performance, revealing potential for synergy. However, the substantial architectural differences between the two fields pose considerable challenges.

Unlike previous approaches, the authors propose UniGenDet: a unified generative-discriminative framework for the co-evolution of image generation and generated-image detection.

Key Contributions

To bridge the task gap, the paper designs a symbiotic multimodal self-attention mechanism and a unified fine-tuning algorithm.
This synergy enables the generation task to improve the interpretability of authenticity recognition, while authenticity criteria guide the creation of higher-fidelity images.
Additionally, a detector-informed generation alignment mechanism is introduced to facilitate seamless information exchange between the two tasks.

Results

Extensive experiments across multiple datasets demonstrate that UniGenDet achieves state-of-the-art performance.

---

*Auto-collected on 2026-04-27*